From: Wu Fengguang <fengguang.wu@intel.com>
To: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Peter Staubach <staubach@redhat.com>,
Myklebust Trond <Trond.Myklebust@netapp.com>,
Jan Kara <jack@suse.cz>,
Andrew Morton <akpm@linux-foundation.org>,
Theodore Tso <tytso@mit.edu>,
Christoph Hellwig <hch@infradead.org>,
Dave Chinner <david@fromorbit.com>,
Chris Mason <chris.mason@oracle.com>,
"Li, Shaohua" <shaohua.li@intel.com>,
"jens.axboe@oracle.com" <jens.axboe@oracle.com>,
Nick Piggin <npiggin@suse.de>,
"linux-fsdevel@vger.kernel.org" <linux-fsdevel@vger.kernel.org>,
Richard Kennedy <richard@rsk.demon.co.uk>,
LKML <linux-kernel@vger.kernel.org>
Subject: Re: [PATCH 01/45] writeback: reduce calls to global_page_state in balance_dirty_pages()
Date: Sat, 17 Oct 2009 13:30:16 +0800 [thread overview]
Message-ID: <20091017053016.GA20086@localhost> (raw)
In-Reply-To: <1255519348.8392.412.camel@twins>
On Wed, Oct 14, 2009 at 07:22:28PM +0800, Peter Zijlstra wrote:
> On Wed, 2009-10-14 at 09:38 +0800, Wu Fengguang wrote:
> > > > Hmm, probably you've discussed this in some other email but why do we
> > > > cycle in this loop until we get below dirty limit? We used to leave the
> > > > loop after writing write_chunk... So the time we spend in
> > > > balance_dirty_pages() is no longer limited, right?
> >
> > Right, this is a legitimate concern.
>
> Quite.
>
> > > Wu was saying that without the loop nr_writeback wasn't limited, but
> > > since bdi_writeback_wakeup() is driven from writeout completion, I'm not
> > > sure how again that was so.
> >
> > Let me summarize the ideas :)
> >
> > There are two cases:
> >
> > - there are no bdi or block io queue to limit nr_writeback
> > This must be fixed. It either let nr_writeback grow to dirty_thresh
> > (with loop) and thus squeeze nr_dirty, or grow out of control
> > totally (without loop). Current state is, the nr_writeback wait
> > queue for NFS is there; the one for btrfs is still missing.
> >
> > - there is a nr_writeback limit, but is larger than dirty_thresh
> > In this case nr_dirty will be close to 0 regardless of the loop.
> > The loop will help to keep
> > nr_dirty + nr_writeback + nr_unstable < dirty_thresh
> > Without the loop, the "real" dirty threshold would be larger
> > (determined by the nr_writeback limit).
> >
> > > We can move all of bdi_dirty to bdi_writeout, if the bdi writeout queue
> > > permits, but it cannot grow beyond the total limit, since we're actually
> > > waiting for writeout completion.
> >
> > Yes, this explains the second case. It's some trade-off like: the
> > nr_writeback limit can not be trusted in small memory systems, so do
> > the loop to impose the dirty_thresh, which unfortunately can hurt
> > responsiveness on all systems with prolonged wait time..
>
> Ok, so I'm still puzzled.
Big sorry - it's me that was confused (by some buggy tests).
> set_page_dirty()
> balance_dirty_pages_ratelimited()
> balance_dirty_pages_ratelimited_nr(1)
> balance_dirty_pages(nr);
>
> So we call balance_dirty_pages() with an appropriate count for each
> set_page_dirty() successful invocation, right?
Right.
> balance_dirty_pages() guarantees that:
>
> nr_dirty + nr_writeback + nr_unstable < dirty_thresh &&
> (nr_dirty + nr_writeback + nr_unstable <
> (dirty_thresh + background_thresh)/2 ||
> bdi_dirty + bdi_writeback + bdi_unstable < bdi_thresh)
>
> Now without loop, without writeback limit, I still see no way to
> actually generate more 'dirty' pages than dirty_thresh.
>
> As soon as we hit dirty_thresh a process will wait for exactly the same
> amount of pages to get cleaned (writeback completed) as were dirtied
> (+/- the ratelimit fuzz which should even out over processes).
Ah yes, we now wait for writeback _completed_ in bdi_writeback_wait(),
instead of _start_ writeback in the old fashioned writeback_inodes().
> That should bound things to dirty_thresh -- the wait is on writeback
> complete, so nr_writeback is bounded too.
Right. It was not bounded in the tests because bdi_writeback_wait()
quits _prematurely_, because the background writeback finds it was
already under background threshold, and so wakeup the throttled tasks
and then quit. Fixed by simply removing the wakeup-all in background
writeback and this change:
if (args->for_background && !over_bground_thresh() &&
+ list_empty(&wb->bdi->throttle_list))
break;
So now
- the throttled tasks are guaranteed to be wakeup
- it will only be wakeup in __bdi_writeout_inc()
- once wakeup, at least write_chunk pages have been written on behalf of it
> [ I forgot the exact semantics of unstable, if we clear writeback before
> unstable, we need to fix something ]
New tests show that NFS is working fine without loop and without NFS
nr_writeback limit:
$ dd if=/dev/zero of=/mnt/test/zero3 bs=1M count=200 &
$ vmmon -d 1 nr_writeback nr_dirty nr_unstable
nr_writeback nr_dirty nr_unstable
0 2 0
0 2 0
0 22477 65
2 20849 1697
2 19153 3393
2 17420 5126
27825 7 5979
27816 0 41
26925 0 907
31064 286 159
32531 0 213
32548 0 89
32405 0 155
32464 0 98
32517 0 45
32560 0 194
32534 0 220
32601 0 222
32490 0 72
32447 0 115
32511 0 48
32535 0 216
32535 0 216
32535 0 216
32535 0 216
31555 0 1180
29732 0 3003
29277 0 3458
27721 0 5014
25955 0 6780
24356 0 8379
22763 0 9972
21083 0 11652
19371 0 13364
17564 0 15171
15781 0 16954
14005 0 18730
12230 0 20505
12177 0 20558
11383 0 21352
9489 0 23246
7621 0 25115
5866 0 26870
4790 0 27947
2962 0 29773
1089 0 31646
0 0 32735
0 0 32735
0 0 0
0 0 0
> Now, a nr_writeback queue that limits writeback will still be useful,
> esp for high speed devices. Once they ramp up and bdi_thresh exceeds the
> queue size, it'll take effect. So you reap the benefits when needed.
Right, the nr_writeback limit avoids
nr_writeback => dirty_thresh
+
nr_dirty + nr_writeback < dirty_thresh
=>
nr_dirty => 0
Thanks for the clarification, it looks less obscure now :)
Thanks,
Fengguang
next prev parent reply other threads:[~2009-10-17 5:30 UTC|newest]
Thread overview: 116+ messages / expand[flat|nested] mbox.gz Atom feed top
2009-10-07 7:38 [PATCH 00/45] some writeback experiments Wu Fengguang
2009-10-07 7:38 ` [PATCH 01/45] writeback: reduce calls to global_page_state in balance_dirty_pages() Wu Fengguang
2009-10-09 15:12 ` Jan Kara
2009-10-09 15:18 ` Peter Zijlstra
2009-10-09 15:47 ` Jan Kara
2009-10-11 2:28 ` Wu Fengguang
2009-10-11 7:44 ` Peter Zijlstra
2009-10-11 10:50 ` Wu Fengguang
2009-10-11 10:58 ` Peter Zijlstra
2009-10-11 11:25 ` Peter Zijlstra
2009-10-12 1:26 ` Wu Fengguang
2009-10-12 9:07 ` Peter Zijlstra
2009-10-12 9:24 ` Wu Fengguang
2009-10-10 21:33 ` Wu Fengguang
2009-10-12 21:18 ` Jan Kara
2009-10-13 3:24 ` Wu Fengguang
2009-10-13 8:41 ` Peter Zijlstra
2009-10-13 18:12 ` Jan Kara
2009-10-13 18:28 ` Peter Zijlstra
2009-10-14 1:38 ` Wu Fengguang
2009-10-14 11:22 ` Peter Zijlstra
2009-10-17 5:30 ` Wu Fengguang [this message]
2009-10-07 7:38 ` [PATCH 02/45] writeback: reduce calculation of bdi dirty thresholds Wu Fengguang
2009-10-07 7:38 ` [PATCH 03/45] ext4: remove unused parameter wbc from __ext4_journalled_writepage() Wu Fengguang
2009-10-07 7:38 ` [PATCH 04/45] writeback: remove unused nonblocking and congestion checks Wu Fengguang
2009-10-09 15:26 ` Jan Kara
2009-10-10 13:47 ` Wu Fengguang
2009-10-07 7:38 ` [PATCH 05/45] writeback: remove the always false bdi_cap_writeback_dirty() test Wu Fengguang
2009-10-07 7:38 ` [PATCH 06/45] writeback: use larger ratelimit when dirty_exceeded Wu Fengguang
2009-10-07 8:53 ` Peter Zijlstra
2009-10-07 9:17 ` Wu Fengguang
2009-10-07 7:38 ` [PATCH 07/45] writeback: dont redirty tail an inode with dirty pages Wu Fengguang
2009-10-09 15:45 ` Jan Kara
2009-10-07 7:38 ` [PATCH 08/45] writeback: quit on wrap for .range_cyclic (write_cache_pages) Wu Fengguang
2009-10-07 7:38 ` [PATCH 09/45] writeback: quit on wrap for .range_cyclic (pohmelfs) Wu Fengguang
2009-10-07 12:32 ` Evgeniy Polyakov
2009-10-07 14:23 ` Wu Fengguang
2009-10-07 7:38 ` [PATCH 10/45] writeback: quit on wrap for .range_cyclic (btrfs) Wu Fengguang
2009-10-07 7:38 ` [PATCH 11/45] writeback: quit on wrap for .range_cyclic (cifs) Wu Fengguang
2009-10-07 7:38 ` [PATCH 12/45] writeback: quit on wrap for .range_cyclic (ext4) Wu Fengguang
2009-10-07 7:38 ` [PATCH 13/45] writeback: quit on wrap for .range_cyclic (gfs2) Wu Fengguang
2009-10-07 7:38 ` [PATCH 14/45] writeback: quit on wrap for .range_cyclic (afs) Wu Fengguang
2009-10-07 7:38 ` [PATCH 15/45] writeback: fix queue_io() ordering Wu Fengguang
2009-10-07 7:38 ` [PATCH 16/45] writeback: merge for_kupdate and !for_kupdate cases Wu Fengguang
2009-10-07 7:38 ` [PATCH 17/45] writeback: only allow two background writeback works Wu Fengguang
2009-10-07 7:38 ` [PATCH 18/45] writeback: introduce wait queue for balance_dirty_pages() Wu Fengguang
2009-10-08 1:01 ` KAMEZAWA Hiroyuki
2009-10-08 1:58 ` Wu Fengguang
2009-10-08 2:40 ` KAMEZAWA Hiroyuki
2009-10-08 4:01 ` Wu Fengguang
2009-10-08 5:59 ` KAMEZAWA Hiroyuki
2009-10-08 6:07 ` Wu Fengguang
2009-10-08 6:28 ` Wu Fengguang
2009-10-08 6:39 ` KAMEZAWA Hiroyuki
2009-10-08 8:08 ` Peter Zijlstra
2009-10-08 8:11 ` KAMEZAWA Hiroyuki
2009-10-08 8:36 ` Jens Axboe
2009-10-09 2:52 ` [PATCH] writeback: account IO throttling wait as iowait Wu Fengguang
2009-10-09 10:41 ` Jens Axboe
2009-10-09 10:58 ` Wu Fengguang
2009-10-09 11:01 ` Jens Axboe
2009-10-08 8:05 ` [PATCH 18/45] writeback: introduce wait queue for balance_dirty_pages() Peter Zijlstra
2009-10-07 7:38 ` [PATCH 19/45] writeback: remove the loop in balance_dirty_pages() Wu Fengguang
2009-10-07 7:38 ` [PATCH 20/45] NFS: introduce writeback wait queue Wu Fengguang
2009-10-07 8:53 ` Peter Zijlstra
2009-10-07 9:07 ` Wu Fengguang
2009-10-07 9:15 ` Peter Zijlstra
2009-10-07 9:19 ` Wu Fengguang
2009-10-07 9:17 ` Nick Piggin
2009-10-07 9:52 ` Wu Fengguang
2009-10-07 7:38 ` [PATCH 21/45] writeback: estimate bdi write bandwidth Wu Fengguang
2009-10-07 8:53 ` Peter Zijlstra
2009-10-07 9:39 ` Wu Fengguang
2009-10-07 7:38 ` [PATCH 22/45] writeback: show bdi write bandwidth in debugfs Wu Fengguang
2009-10-07 7:38 ` [PATCH 23/45] writeback: kill space in debugfs item name Wu Fengguang
2009-10-07 7:38 ` [PATCH 24/45] writeback: remove global nr_to_write and use timeout instead Wu Fengguang
2009-10-07 7:38 ` [PATCH 25/45] writeback: convert wbc.nr_to_write to per-file parameter Wu Fengguang
2009-10-07 7:38 ` [PATCH 26/45] block: pass the non-rotational queue flag to backing_dev_info Wu Fengguang
2009-10-07 7:38 ` [PATCH 27/45] writeback: introduce wbc.for_background Wu Fengguang
2009-10-07 7:38 ` [PATCH 28/45] writeback: introduce wbc.nr_segments Wu Fengguang
2009-10-07 7:38 ` [PATCH 29/45] writeback: fix the shmem AOP_WRITEPAGE_ACTIVATE case Wu Fengguang
2009-10-07 11:57 ` Hugh Dickins
2009-10-07 14:00 ` Wu Fengguang
2009-10-07 7:38 ` [PATCH 30/45] vmscan: lumpy pageout Wu Fengguang
2009-10-07 7:38 ` [PATCH 31/45] writeback: sync old inodes first in background writeback Wu Fengguang
2010-07-12 3:01 ` Christoph Hellwig
2010-07-12 15:24 ` Wu Fengguang
2009-10-07 7:38 ` [PATCH 32/45] writeback: update kupdate expire timestamp on each scan of b_io Wu Fengguang
2009-10-07 7:38 ` [PATCH 34/45] writeback: sync livelock - kick background writeback Wu Fengguang
2009-10-07 7:38 ` [PATCH 35/45] writeback: sync livelock - use single timestamp for whole sync work Wu Fengguang
2009-10-07 7:38 ` [PATCH 36/45] writeback: sync livelock - curb dirty speed for inodes to be synced Wu Fengguang
2009-10-07 7:38 ` [PATCH 37/45] writeback: use timestamp to indicate dirty exceeded Wu Fengguang
2009-10-07 7:38 ` [PATCH 38/45] writeback: introduce queue b_more_io_wait Wu Fengguang
2009-10-07 7:38 ` [PATCH 39/45] writeback: remove wbc.more_io Wu Fengguang
2009-10-07 7:38 ` [PATCH 40/45] writeback: requeue_io_wait() on I_SYNC locked inode Wu Fengguang
2009-10-07 7:38 ` [PATCH 41/45] writeback: requeue_io_wait() on pages_skipped inode Wu Fengguang
2009-10-07 7:39 ` [PATCH 42/45] writeback: requeue_io_wait() on blocked inode Wu Fengguang
2009-10-07 7:39 ` [PATCH 43/45] writeback: requeue_io_wait() on fs redirtied inode Wu Fengguang
2009-10-07 7:39 ` [PATCH 44/45] NFS: remove NFS_INO_FLUSHING lock Wu Fengguang
2009-10-07 13:11 ` Peter Staubach
2009-10-07 13:32 ` Wu Fengguang
2009-10-07 13:59 ` Peter Staubach
2009-10-08 1:44 ` Wu Fengguang
2009-10-07 7:39 ` [PATCH 45/45] btrfs: fix race on syncing the btree inode Wu Fengguang
2009-10-07 8:53 ` [PATCH 00/45] some writeback experiments Peter Zijlstra
2009-10-07 10:17 ` [PATCH 14/45] writeback: quit on wrap for .range_cyclic (afs) David Howells
2009-10-07 10:21 ` Nick Piggin
2009-10-07 10:47 ` Wu Fengguang
2009-10-07 11:23 ` Nick Piggin
2009-10-07 12:21 ` Wu Fengguang
2009-10-07 13:47 ` [PATCH 00/45] some writeback experiments Peter Staubach
2009-10-07 15:18 ` Wu Fengguang
2009-10-08 5:33 ` Wu Fengguang
2009-10-08 5:44 ` Wu Fengguang
2009-10-07 14:26 ` Theodore Tso
2009-10-07 14:45 ` Wu Fengguang
Reply instructions:
You may reply publicly to this message via plain-text email
using any one of the following methods:
* Save the following mbox file, import it into your mail client,
and reply-to-all from there: mbox
Avoid top-posting and favor interleaved quoting:
https://en.wikipedia.org/wiki/Posting_style#Interleaved_style
* Reply using the --to, --cc, and --in-reply-to
switches of git-send-email(1):
git send-email \
--in-reply-to=20091017053016.GA20086@localhost \
--to=fengguang.wu@intel.com \
--cc=Trond.Myklebust@netapp.com \
--cc=a.p.zijlstra@chello.nl \
--cc=akpm@linux-foundation.org \
--cc=chris.mason@oracle.com \
--cc=david@fromorbit.com \
--cc=hch@infradead.org \
--cc=jack@suse.cz \
--cc=jens.axboe@oracle.com \
--cc=linux-fsdevel@vger.kernel.org \
--cc=linux-kernel@vger.kernel.org \
--cc=npiggin@suse.de \
--cc=richard@rsk.demon.co.uk \
--cc=shaohua.li@intel.com \
--cc=staubach@redhat.com \
--cc=tytso@mit.edu \
/path/to/YOUR_REPLY
https://kernel.org/pub/software/scm/git/docs/git-send-email.html
* If your mail client supports setting the In-Reply-To header
via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line
before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).